On 25/06/21 09:28, Peter Zijlstra wrote:

On Fri, Jun 25, 2021 at 11:16:08AM +0530, Srikar Dronamraju wrote:

quoted

Bharata,

I think the regression is due to Commit f1a0a376ca0c ("sched/core:
Initialize the idle task with preemption disabled")

So that extra preempt_disable() that got removed would've incremented it
to 1 and then things would've been fine.

Except.. Valentin changed things such that preempt_count() should've
been inittialized to 1, instead of 0, but for some raisin that didn't
stick.. what gives.

So we have init_idle(p) -> init_idle_preempt_count(p) ->
task_thread_info(p)->preempt_count = PREEMPT_DISABLED;

But somehow, by the time you're running start_secondary(), that's gotten
to be 0 again. Does DEBUG_PREEMPT give more clues?

Given the preempt_count isn't reset between hotplugs anymore, you might be
able to find the culprit with a hotplug cycle and ftrace with
trace_prempt_off and trace_preempt_on events (requires PREEMPT_TRACER
IIRC).

It's doable at boot time too, but that will mean sifting through many more
events than you'd like...